Re: Recent removal of bsg read/write support

[Michal Hocko]

On Tue 04-09-18 05:38:21, Douglas Gilbert wrote:
> On 2018-09-03 02:28 PM, Michal Hocko wrote:
> > On Sun 02-09-18 21:16:10, Douglas Gilbert wrote:
> > > On 2018-09-02 01:44 PM, Richard Weinberger wrote:
> > > > CC'ing relevant people. Otherwise your mail might get lost.
> > > > 
> > > > On Sun, Sep 2, 2018 at 1:37 PM Dror Levin <drorl@xxxxxxxxxxxxx> wrote:
> > > > > 
> > > > > Note: I am not subscribed to LKML so please CC replies to this email.
> > > > > 
> > > > > Hi,
> > > > > 
> > > > > We have an internal tool that uses the bsg read/write interface to
> > > > > issue SCSI commands as part of a test suite for a storage device.
> > > > > 
> > > > > After recently reading on LWN that this interface is to be removed we
> > > > > tried porting our code to use sg instead. However, that raises new
> > > > > issues - mainly getting ENOMEM over iSCSI for unknown reasons.
> > > > > 
> > > > > Because of this we would like to continue using the bsg interface,
> > > > > even if some changes are required to meet security concerns.
> > > > > 
> > > > > Is there any chance for this removal to be reverted? I saw it was
> > > > > already included in 4.19-rc1.
> > > 
> > > Hi,
> > > Both bsg and sg are relatively thin shims over the same block layer
> > > pass-through calls. And neither driver will themselves generate ENOMEM
> > > unless the CPU is running low of memory.
> > > 
> > > In my experience, the main reason for unexpected ENOMEMs *** is from
> > > blk_rq_map_user_iov() in block/blk_map.c called from both drivers.
> > > That is a particular resource shortage rather than memory in general.
> > > I do notice the blk_rq_map_user_iov() is/was called with GFP_KERNEL
> > > in bsg and GFP_ATOMIC by sg. That suggests when you call write() on
> > > a sg device and get ENOMEM, then wait a little (depends on your app)
> > > and try again.
> > 
> > Well, what is the reason to use GFP_ATOMIC in the first place? I am not
> > familiar with the code so I might be easily wrong but sg_start_req which
> > calls blk_rq_map_user_iov resp. blk_rq_map_user with GFP_ATOMIC uses
> > mutex. It is a conditional usage so the sleeping context might depend
> > on the caller. But I guess it would be better to double check. It looks
> > suspicious to me.
> 
> Of the hundreds of 'hacks' on the sg driver over the years, the most
> common is an expert arguing that GFP_ATOMIC should be changed to GFP_KERNEL.
> They usually get their way. That is followed around 6 to 9 months later by
> a sg user complaining about an unexpected broken app. So back it goes to
> GFP_ATOMIC.

Then I would strongly recommend to describe the actual reTquirements on
the allocation context. Why is GFP_ATOMIC really needed? There are
usually two reasons a) the allocation is called from an atomic context
b) reclaim is not acceptable or desirable.
-- 
Michal Hocko
SUSE Labs